FCG-DETR: Foreground Confidence-Guided Detection Transformer for Remote Sensing Oriented Object Detection

Accurate detection of oriented objects in remote sensing imagery is often hindered by various aspect ratios and complex background interference. While Transformer based detectors establish an end-to-end pipeline by eliminating non-maximum suppression (NMS), they incur substantial computational overhead due to indiscriminate processing of all tokens. This limitation not only leads to unnecessary resource consumption but also compromises detection performance. To address this issue, we propose FCG-DETR, an efficient oriented object detection framework that leverages hierarchical foreground guidance, dynamic token selection, and confidence-gated decoding to reduce computational overhead while improving detection accuracy. Specifically, the Hierarchical Foreground Generation (HFG) module leverages anisotropic Gaussian modeling to generate category-agnostic spatial priors, which serve as soft-supervision signals to guide the encoder toward high potential regions across varying scales and aspect ratios.To further optimize efficiency, we design a Dynamic Foreground Token Selection (DFS) strategy. Instead of uniform sampling, DFS utilizes normalized cross scale metrics to adaptively align token distribution with the object density of each feature level. Finally, a Foreground Confidence-Gated Attention (FCGA) mechanism is proposed to resolve the optimization conflict in the decoder. By dynamically modulating the reliance on spatial priors based on query confidence, FCGA enables precise refinement for converged queries while ensuring robust exploration for uncertain ones. Extensive experiments on DOTA and DIOR-R benchmarks demonstrate that FCG-DETR achieves state-of-the-art performance with faster convergence and significantly reduced computational overhead. The code is available at https://github.com/gyy1502/FCG-DETR.